def __init__( self, data_root: str, csv: str, split: str, tokenizer: SentencePieceBPETokenizer, image_transform: Callable = T.DEFAULT_IMAGE_TRANSFORM, padded_length: int = 256, max_caption_length: int = 50, use_single_caption: bool = False, percentage: float = 100.0, ): self.data_root = data_root self.padded_length = padded_length info_df = pd.read_csv(os.path.join(data_root, csv), delimiter="|") self.video_list = [] for index, row in info_df.iterrows(): self.video_list.append((index, row['name'], [row['orth']])) self.image_transform = image_transform self.caption_transform = alb.Compose([ T.NormalizeCaption(), T.TokenizeCaption(tokenizer), T.TruncateCaptionTokens(max_caption_length), ]) self.use_single_caption = use_single_caption self.padding_idx = tokenizer.token_to_id("<unk>")
def __init__( self, data_root: str, split: str, tokenizer: SentencePieceBPETokenizer, image_transform: Callable = T.DEFAULT_IMAGE_TRANSFORM, mask_proportion: float = 0.15, mask_probability: float = 0.80, replace_probability: float = 0.10, max_caption_length: int = 30, use_single_caption: bool = False, percentage: float = 100.0, ): lmdb_path = os.path.join(data_root, 'virtex', f"serialized_{split}.lmdb") self.reader = LmdbReader(lmdb_path, percentage=percentage) self.image_transform = image_transform self.caption_transform = alb.Compose([ T.NormalizeCaption(), T.TokenizeCaption(tokenizer), T.TruncateCaptionTokens(max_caption_length), ]) self.use_single_caption = use_single_caption self.padding_idx = tokenizer.token_to_id("<unk>") # Handles to commonly used variables for word masking. self._vocab_size = tokenizer.get_vocab_size() self._mask_index = tokenizer.token_to_id("[MASK]") self._mask_proportion = mask_proportion self._mask_prob = mask_probability self._repl_prob = replace_probability
def __init__( self, data_root: str, split: str, tokenizer: SentencePieceBPETokenizer, image_transform: Callable = T.DEFAULT_IMAGE_TRANSFORM, max_caption_length: int = 30, ): lmdb_path = os.path.join(data_root, f"serialized_{split}.lmdb") self.reader = LmdbReader(lmdb_path) self.image_transform = image_transform self.caption_transform = alb.Compose([ T.NormalizeCaption(), T.TokenizeCaption(tokenizer), T.TruncateCaptionTokens(max_caption_length), ]) self.padding_idx = tokenizer.token_to_id("<unk>")
def __init__(self, data_root: str, split: str, tokenizer: SentencePieceBPETokenizer, image_transform: Callable = T.DEFAULT_IMAGE_TRANSFORM, max_caption_length: int = 77, percentage: float = 100.0, all_captions: bool = False, include_image: bool = True): super().__init__() self.percentage = percentage self.tokenizer = tokenizer self.image_transform = image_transform self.caption_transform = alb.Compose([ T.NormalizeCaption(), T.TokenizeCaption(tokenizer), T.TruncateCaptionTokens(max_caption_length), ]) self.padding_idx = tokenizer.pad_id self.data_root = data_root self.split = split self._all_captions = all_captions self._include_image = include_image